Instructions to use DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForMultimodalLM

tokenizer = AutoTokenizer.from_pretrained("DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark")
model = AutoModelForMultimodalLM.from_pretrained("DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark

SGLang

How to use DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark with Docker Model Runner:
```
docker model run hf.co/DreamFast/GLM-4.7-Flash-Uncensored-HauhauCS-Safetensor-Benchmark
```

DreamFast commited on Apr 25

Commit

5bdce2a

verified ·

1 Parent(s): 5d16c3d

Add Abliterlitics repo link

Browse files

Files changed (1) hide show

README.md +3 -1

README.md CHANGED Viewed

@@ -16,6 +16,8 @@ tags:
 # GLM-4.7-Flash: HauhauCS, Safetensors
 This is the HauhauCS abliteration of [GLM-4.7-Flash](https://huggingface.co/zai-org/GLM-4.7-Flash), converted from the BF16 GGUF release to native safetensors using [ungguf](https://github.com/dreamfast/ungguf).
 HauhauCS claims these are *"No changes to datasets or capabilities. Fully functional, 100% of what the original authors intended, just without the refusals"* and describes them as *"the best lossless uncensored models out there."*
@@ -304,7 +306,7 @@ Weight forensics reveal that HauhauCS used four stacked methods from the reaper-
 - **Safety:** [HarmBench](https://github.com/centerforaisafety/HarmBench) 400 textual behaviours, `max_tokens=2048, temperature=0.0`, `classify_harmbench.py` v3.0 with manual overrides, reviewed by GLM 5.1
 - **KL divergence:** `F.kl_div(logprobs_variant, logprobs_base, reduction="batchmean", log_target=True)` on full vocab first-token logits via `model.generate(max_new_tokens=1, output_scores=True)`, matching the [Heretic evaluator](https://github.com/p-e-w/heretic/blob/master/src/heretic/evaluator.py). Dataset: [mlabonne/harmless_alpaca](https://huggingface.co/datasets/mlabonne/harmless_alpaca) `test[:100]`, system prompt "You are a helpful assistant." Collected with BF16 dual-GPU inference (RTX 5090 + RTX 4090) with CPU offloading. Validated on single A100-80GB (no offload) for Heretic: KL=0.0115 vs 0.0110, confirming offload does not meaningfully distort results.
 - **CoT forensics:** Keyword-based analysis of 2,000 HarmBench reasoning chains (400 per model) captured via OpenAI-compatible API `reasoning` field. Patterns detected: safety deliberation, explicit refusal language, educational pivots, disclaimers.
-- **Weight analysis:** SVD, fingerprint, edit vector overlap, per-layer analysis, rank structure, and cross-technique alignment comparing all four abliteration variants against the base
 - **Hardware:** RTX 5090 32GB + RTX 4090 24GB
 ## Forensic Notes

 # GLM-4.7-Flash: HauhauCS, Safetensors
+> Forensic analysis by [Abliterlitics](https://github.com/dreamfast/abliterlitics) — open-source abliteration forensics toolkit
 This is the HauhauCS abliteration of [GLM-4.7-Flash](https://huggingface.co/zai-org/GLM-4.7-Flash), converted from the BF16 GGUF release to native safetensors using [ungguf](https://github.com/dreamfast/ungguf).
 HauhauCS claims these are *"No changes to datasets or capabilities. Fully functional, 100% of what the original authors intended, just without the refusals"* and describes them as *"the best lossless uncensored models out there."*
 - **Safety:** [HarmBench](https://github.com/centerforaisafety/HarmBench) 400 textual behaviours, `max_tokens=2048, temperature=0.0`, `classify_harmbench.py` v3.0 with manual overrides, reviewed by GLM 5.1
 - **KL divergence:** `F.kl_div(logprobs_variant, logprobs_base, reduction="batchmean", log_target=True)` on full vocab first-token logits via `model.generate(max_new_tokens=1, output_scores=True)`, matching the [Heretic evaluator](https://github.com/p-e-w/heretic/blob/master/src/heretic/evaluator.py). Dataset: [mlabonne/harmless_alpaca](https://huggingface.co/datasets/mlabonne/harmless_alpaca) `test[:100]`, system prompt "You are a helpful assistant." Collected with BF16 dual-GPU inference (RTX 5090 + RTX 4090) with CPU offloading. Validated on single A100-80GB (no offload) for Heretic: KL=0.0115 vs 0.0110, confirming offload does not meaningfully distort results.
 - **CoT forensics:** Keyword-based analysis of 2,000 HarmBench reasoning chains (400 per model) captured via OpenAI-compatible API `reasoning` field. Patterns detected: safety deliberation, explicit refusal language, educational pivots, disclaimers.
+- **Weight analysis:** SVD, fingerprint, edit vector overlap, per-layer analysis, rank structure, and cross-technique alignment comparing all four abliteration variants against the base, using [Abliterlitics](https://github.com/dreamfast/abliterlitics)
 - **Hardware:** RTX 5090 32GB + RTX 4090 24GB
 ## Forensic Notes